|Publication number||US7487279 B2|
|Application number||US 11/625,842|
|Publication date||Feb 3, 2009|
|Filing date||Jan 23, 2007|
|Priority date||Jan 23, 2007|
|Also published as||EP2126704A2, EP2126704A4, EP2126704B1, US20080177955, WO2008091757A2, WO2008091757A3|
|Publication number||11625842, 625842, US 7487279 B2, US 7487279B2, US-B2-7487279, US7487279 B2, US7487279B2|
|Original Assignee||International Business Machines Corporation|
|Export Citation||BiBTeX, EndNote, RefMan|
|Patent Citations (6), Non-Patent Citations (5), Referenced by (5), Classifications (10), Legal Events (5)|
|External Links: USPTO, USPTO Assignment, Espacenet|
The invention disclosed broadly relates to the field of optimizing lock operations and more particularly relates to the field of spin locks.
Computer systems can suffer severe performance degradation as a result of high lock contention. Lock operations are associated with software locks which are used to ensure that only one process at a time can access a shared resource, such as a region of memory. The act of accessing the shared resource is also referred to as entering the critical region. Most shared-memory architectures, such as SMP (Symmetric Multi-Processor) and NUMA (Non-Uniform Memory Access) architectures also provide hardware support for making mutually exclusive accesses to shared cache data. This hardware support is known as a cache coherent mechanism. Locks are fundamental software synchronization primitives for controlling concurrent access to shared data. The performance degradation occurs under high lock contention because only one CPU can acquire the lock and do useful work while all other CPUs must wait for the lock to be released.
Spin locks are a very simple and, in the right context, efficient method of synchronizing access to shared resources. A “spin lock” is so named because if a spin lock is not available, the caller busy-waits (or “spins”) until the lock becomes available. This is called “software spin-waiting.” The algorithm for spin-waiting is very simple. Each process must check a shared lock variable. When a shared resource is required, a process calls lock( ) to test and set the variable and when the resource is ready to be released, a process calls unlock( ) to clear the variable. The lock( ) function will cause a waiting process to loop until the resource is available. The availability of the resource is defined by the value of the shared lock variable. Conventionally, if the value of this variable is zero, then the resource is available, otherwise it is in use by another process. When the resource is locked, the shared variable holds a non-zero value.
Systems employing locks typically require that a given process perform an atomic operation to obtain access to a shared data structure. In other words, another process cannot access the lock between the test and set portion of the atomic operation. The test-and-set operation is performed to determine whether a lock variable associated with the shared data structure is cleared and to atomically set the lock variable. That is, the test allows the process to determine whether the shared data structure is free of a lock by another process, and the set operation allows the process to acquire the lock if the lock variable is cleared.
The key to lock design is to achieve both fairness at high contention and performance at low contention. Fairness in the context of lock design is defined as equalizing the ability of all contenders vying for a lock to acquire that lock, regardless of the contenders' position and/or relationship to the existing lock holder. Existing locking mechanisms have achieved fairness or high performance, but not both. Specifically, three of the most widely used lock designs are: spin locks, queue locks, adaptive locks, and fairlocks.
Spin locks are widely used due to their high performance at low contention. [See T. E. Anderson., The Performance Implications of Spin Lock Alternatives for Shared-Memory Multiprocessors. IEEE Transaction on Parallel and Distributed Systems, 1(1):6-16, January 1990] In a spin lock implementation, if the test of the lock variable indicates that another process has acquired the lock to the shared memory region, the requesters for the lock initiate a loop wherein the lock variable is continuously read until the lock variable is cleared, at which time the waiting processes reinitiate the atomic test-and-set operation. In NUMA architectures, spin locks can create locking starvation due to the locking unfairness created by the large memory access latency difference across nodes. Because a CPU can access a spin lock in its local cache much faster than a CPU can access the same spin lock from another node. Therefore, the CPU that is on the same node where the spin lock is located has a much higher chance of acquiring the lock than a CPU that resides on another node.
Queue locks [see J. M. Mellor-Crummey and M. L. Scott, Algorithms for Scalable Synchronization on Shared-Memory Multiprocessors. ACM Transactions on Computer Systems, 9(1):21-65, February 1991] and adaptive locks [see B.-H. Lim and A. Agarwal, Reactive Synchronization Algorithms for Multiprocessors. ASPLOS 1994] avoid unfairness under high contention by introducing complex lock data structures and algorithms to keep track of the usage and ownership of the lock. However, these data structures and algorithms also introduce additional locking latency overhead, therefore sacrificing locking performance at low contention.
Fairlocks provide fairness at high contention by using a bit mask to keep track of contending CPUs and enforce fairness by asking the lock-releasing CPU to explicitly yield the lock to other contending CPUs. [see S. Swaminathan, J. Stultz, J. F. Vogel, P. McKenney, Fairlocks—A High Performance Fair Locking Scheme. 14th International Conference on Parallel and Distributed Computing and Systems, November 2002] Due to the simpler data structure and algorithm used, fairlocks have better locking performance than queue locks and adaptive locks at low contention. But their locking performance at low contention is still worse than that of spin locks.
There is a need for an improved lock mechanism that overcomes the shortcomings of the prior art.
Briefly, according to an embodiment of the invention, a method of implementing a spin lock in a system including a plurality of processing nodes, each node including at least one processor and a cache memory, includes steps of: acquiring exclusivity to the cache memory; checking the availability of the spin lock; setting the spin lock to logical one if the spin lock is available; then setting the spin lock to logical zero once processing is complete; and explicitly yielding the cache memory exclusivity. Yielding the cache memory exclusivity includes instructing the cache coherent hardware to mark the cache memory as non-exclusive. The cache memory, typically called a Level 2 (L2) cache, is physically divided among multiple processing nodes but logically shared as a single addressable unit among the multiple processing nodes.
The features believed characteristic of the invention are set forth in the appended claims. The invention itself, however, as well as an example of its preferred mode of use, further objectives and advantages thereof, will best be understood by reference to the following detailed description of an illustrative embodiment when read in conjunction with the accompanying drawings, wherein:
While the invention as claimed can be modified into alternative forms, specific embodiments thereof are shown by way of example in the drawings and will herein be described in detail. It should be understood, however, that the drawings and detailed description thereto are not intended to limit the invention to the particular form disclosed, but on the contrary, the intention is to cover all modifications, equivalents and alternatives falling within the scope of the present invention.
We describe a method for providing locking fairness at high contention without sacrificing locking performance at low contention for large-scale, Non-Uniform Memory Access (NUMA) distributed shared memory multiprocessor architectures, according to an embodiment of the invention. This is achieved through a key aspect of the invention:
A “cooperative” feature to the existing spin lock mechanism provides statistical locking fairness at high contention without any additional lock data structure and algorithm; therefore retaining its locking performance at low contention. It is more beneficial to provide statistical locking fairness rather than the absolute locking fairness targeted by existing locking mechanisms. Statistical fairness means that all of the lock contenders have an equal chance of acquiring a lock, while absolute fairness means that all of the lock contenders have an equal number of times a lock is actually acquired. While statistical fairness is a weaker condition than absolute fairness, statistical fairness is adequate for applications in all practical purposes and will optimize performance by reducing locking overhead. Statistical fairness is considered a weaker condition because it does not guarantee that at any given instance of time each lock contender acquires the lock exactly the same number of times. Rather, it only provides that each lock contender has an equal chance of acquiring the lock. Therefore, at any given instance of time, when one looks at how many times each lock contender has acquired the lock, the numbers may not be all the same. But as time goes to infinite, each lock contender will acquire the lock the same number of times.
The present invention addresses the problem of providing both locking performance and locking fairness at the same time in Non-Uniform Memory Access (NUMA) distributed shared memory multiprocessor architectures. The “cooperative” feature to the existing spin lock mechanism enables locking fairness under high contention without sacrificing locking performance under low contention. The cooperative feature does not require any change to the current spin lock usage interface and semantics; therefore it does not require any change to the applications utilizing spin locks either.
A lock recoverability mechanism can be used for those uncommon situations where the processing aborts before the lock is released in order to make it available to another process.
For simplicity, a memory address is used as an example of a resource associated with a lock and a CPU is used as an example of a device seeking the resource. It should be noted that other devices and other resources, such as input/output interfaces, where locks are applicable, can also benefit from the embodiment according to the present invention and fall within the scope of the invention.
Referring now in specific detail to the drawings, and particularly to
According to another embodiment of the invention, a computer readable medium, such as a CDROM or DVDROM 190 can include program instructions for operating a portion of the system 100 according to the invention. What has been shown and discussed is a highly-simplified depiction of a programmable computer apparatus. Those skilled in the art will appreciate that other low-level components and connections are required in any practical application of a computer apparatus.
Referring now to
With reference now to
When a CPU seeks a data item from memory it follows a logical progression of searching for the data item in the memory hierarchy, with access latency increasing progressively. First, the CPU determines if the data item is available in its private Level 1 (L1) Cache (not shown in the figure since it is irrelevant to this invention). If the data item is not available in the private L1 Cache, the local L2 Cache is checked. If the data item is not available in the local L2 Cache, the L2 Caches of all other nodes within the same network are checked. Finally, the data item is fetched from the Main Memory 150 if it is not available from any of the L2 Caches. The important memory access characteristics to bear in mind are: a) accessing local L2 Cache is typically the fastest (again, ignoring L1 Cache), most direct path to the desired item; b) accessing remote L2 Cache is typically two to three times slower than accessing local L2 Cache; and c) accessing Main Memory 150 is typically two to three times slower than accessing remote L2 Cache.
To understand why current spin locks create locking unfairness in NUMA architectures as depicted in
Once the cache exclusivity is obtained, the spin lock 155 is tested to check its value in step 310. If the spin lock 155 is not equal to zero, this indicates that the spin lock 155 is locked by another processor. Since the CPU cannot acquire a spin lock with a non-zero value, the processing loops back to step 305 and the CPU tries to acquire the spin lock 155 again. This loop will continue until the spin lock 155 has been released by the holding CPU. Releasing a spin lock is accomplished by writing the value of “zero” into the spin lock. Once the value has been set to zero, this indicates that the holding CPU has released the spin lock 155. It is now therefore available for the acquiring CPU to write the value “one” into the spin lock 155.
If the spin lock 155 is equal to zero, it is now available to another processor. At this point the processing moves on to step 315 and a logical one is stored into the spin lock 155. Setting the spin lock 155 to logical one is an indicator that the spin lock 155 has been acquired. This makes it unavailable to another process. A copy 145 of the spin lock 155 is now brought into the L2 Cache from Main Memory 150 and the acquiring CPU now holds exclusivity to the L2 Cache. Next in step 320 the CPU can proceed to perform normal data processing routines, such as reading and/or writing whatever shared data it needed, and so forth, while it has control of the spin lock 145. Note that setting a spin lock to zero or one is an example of switching a lock “on” or “off.”
Once the CPU is finished its processing, it will prepare to release the spin lock 145. In order to release the spin lock 145, in step 325 the CPU must again obtain exclusivity to the L2 Cache because a value zero must be stored into the spin lock 145 in order to release it. The holding CPU may have temporarily lost L2 Cache exclusivity because another processor may have gained exclusivity while attempting to acquire the spin lock 145. The CPU had previously held exclusivity to the L2 Cache but another CPU that is trying to acquire the spin lock (but will fail since the first CPU is still holding it) “stole” the L2 exclusivity from the first CPU. Note that acquiring a software spin lock and obtaining hardware cache exclusivity are two different processes. In order to acquire a software spin lock, i.e., store one into the spin lock, a CPU must obtain hardware cache exclusivity to the spin lock. However, having obtained hardware cache exclusivity to the spin lock does not mean that the CPU has acquired the software spin lock, unless the spin lock has value zero, i.e., it is not held by another CPU. A CPU cannot write to a spin lock without having exclusive rights to the cache where the spin lock is located. Note that two levels of exclusivity must be breached in order to have access to data. Once cache exclusivity is regained, the lock 145 is set to a logical zero in step 330. Setting the spin lock 145 to zero essentially releases it, making it available to a waiting process.
It is important to note that for both acquiring and releasing the spin lock 145, the hardware must first obtain exclusivity to the spin lock 145 in L2 Cache (330 and 335) because a value must be stored into the spin lock 145. And below we will see that it is the necessity of obtaining exclusivity to the spin lock 145 in L2 Cache for both acquiring (step 205) and releasing (step 330) the spin lock 145 that results in the locking unfairness.
We now use an example to illustrate how locking unfairness can happen, referring to both
During this time, the other three CPUs: CPU2 120, CPU3 225, and CPU4 230, will continue their attempts at acquiring the spin lock 155 by “spin waiting;” and either the L2 Cache1 135 (if CPU2 attempts to acquire the spin lock) or the L2 Cache2 240 (if CPU3 or CPU4 attempt to acquire the spin lock) may take turns to hold exclusivity to the spin lock 155. However, none of the three CPUs can acquire the spin-lock 155 since the value of the spin lock will remain at logical one before CPU1 115 releases the spin lock 155.
After CPU1 115 finishes its data processing and releases the spin lock 155 in step 330, its associated L2 Cache1 135 must first re-obtain exclusivity to the spin lock 155 in order to turn the spin lock “off” (write a zero into it). After CPU 115 obtains exclusivity and releases the spin lock 155, it does not voluntarily give up its exclusivity to the L2 Cache. This is where the unfairness is evidenced. Remember that the spin lock 155 is currently residing in L2 Cache1 135 of Node1 105 even though it has been released. This is because CPU1 115 had to re-obtain exclusivity to its cache. And keep in mind that accessing a local cache is faster than accessing a remote cache to obtain exclusivity. Therefore, regardless of which node's L2 Cache (135 or 240) has been holding exclusivity to the spin lock 145 in its attempt to acquire it during the time when CPU1 115 held the spin lock 145, now the exclusivity is back with L2 Cache1 135. Subsequently, CPU2 120 will have a higher chance of acquiring the just released spin-lock 155 than either CPU3 225 or CPU4 230. This is because the step for obtaining exclusivity to the spin lock 145 in L2 Cache (step 205) which is necessary for acquiring the spin lock 155, completes much faster for CPU2 120 than for CPU3 or CPU4 (225 or 230) because the L2 Cache1 135 already holds exclusivity to the spin lock 145 due to the releasing of the spin lock 155 by CPU1 115. On the other hand, for the L2 Cache2 240 to acquire exclusivity to the spin lock 145, it needs to conduct inter-node cache interrogation through Cache Coherent Hardware 180 (represented by dashed line 260), which is typically two to three times slower than acquiring exclusivity within the same L2 Cache.
The current method of spin lock exclusivity confers an advantage to the co-processors within the same node because of the current practice of a CPU not voluntarily giving up its cache exclusivity after releasing a lock.
Now that it is understood why current spin locks create locking unfairness under high contention in NUMA architectures, we explain how to remove the locking unfairness using the “cooperative” feature of the present invention to resolve the problem. Refer to
The cooperative feature comes into play in step 440. After the logical zero is stored into the spin lock 155 (in step 330), the L2 Cache1 135 explicitly yields its exclusivity to the spin lock 155 by telling the cache coherent hardware to mark the L2 Cache1 135 as non-exclusive to any CPU, even though technically CPU2 115 has exclusivity to the cache. Consequently, CPU2 120 no longer has an advantage because the Level 2 cache resident in its node already holds exclusivity to the spin lock 145. Therefore, both the L2 Cache1 135 and the L2 Cache2 140 must conduct inter-cache interrogation 160 to obtain exclusivity to the spin lock 155. As a result, the other three CPUs, CPU2 120, CPU3 225, and CPU4 230, will be competing equally for acquiring the spin lock 155 just released by CPU1 115 and statistical locking fairness under high contention is achieved.
Since the cooperative feature does not require any additional lock data structure and algorithm, locking performance under low contention is preserved. In addition, no change to the current spin lock usage interface and semantics is required; therefore no change to the applications utilizing spin lock is required either.
Therefore, while there has been described what is presently considered to be the preferred embodiment, it will understood by those skilled in the art that other modifications can be made within the spirit of the invention.
|Cited Patent||Filing date||Publication date||Applicant||Title|
|US5442763 *||Dec 6, 1994||Aug 15, 1995||International Business Machines Corporation||System and method for preventing deadlock in multiprocessor multiple resource instructions|
|US6260117 *||Jan 8, 1998||Jul 10, 2001||International Business Machines Corporation||Method for increasing efficiency in a multi-processor system and multi-processor system with increased efficiency|
|US6275907 *||Nov 2, 1998||Aug 14, 2001||International Business Machines Corporation||Reservation management in a non-uniform memory access (NUMA) data processing system|
|US6779090 *||May 21, 2002||Aug 17, 2004||International Business Machines Corporation||Spinlock for shared memory|
|US6792497 *||May 9, 2001||Sep 14, 2004||Unisys Corporation||System and method for hardware assisted spinlock|
|US7246187 *||Sep 30, 2004||Jul 17, 2007||Emc Corporation||Method and apparatus for controlling exclusive access to a shared resource in a data storage system|
|1||B.-H. Lim and A. Agarwal, Reactive Synchronization Algorithms for Multiprocessors. ASPLOS 1994.|
|2||*||Definition of Cache Coherency from Wikipedia, undated.|
|3||J.M. Mellor-Crummey and M.L. Scott, Algorithms for Scalable Synchronization on Shared-Memory Multiprocessors. ACM Transactions on Computer Systems, 9(1):21-65, Feb. 1991.|
|4||S. Swaminathan, J. Stultz, J.F. Vogel, P. McKenney, Fairlocks-A High Performance Fair Locking Scheme. 14th International Conference on Parallel and Distributed Computing and Systems, Nov. 2002.|
|5||T.E. Anderson, The Performance Implications of Spin Lock Alternatives for Shared-Memory Multiprocessors. IEEE Transaction on Parallel and Distributed Systems, 1(1):6-16, Jan. 1990.|
|Citing Patent||Filing date||Publication date||Applicant||Title|
|US8549249 *||Sep 20, 2010||Oct 1, 2013||Tilera Corporation||Supporting secondary atomic operations using primary atomic operations|
|US8694706||Apr 27, 2012||Apr 8, 2014||Oracle International Corporation||System and method for NUMA-aware locking using lock cohorts|
|US8966491||Apr 27, 2012||Feb 24, 2015||Oracle International Corporation||System and method for implementing NUMA-aware reader-writer locks|
|US9189228 *||Jul 1, 2011||Nov 17, 2015||Adobe Systems Incorporated||Systems and methods for developing and testing electronic content using multiple devices|
|US9697055||Sep 22, 2015||Jul 4, 2017||International Business Machines Corporation||Almost fair busy lock|
|U.S. Classification||710/200, 710/244, 710/240|
|Cooperative Classification||G06F12/0828, G06F12/0811, G06F9/526|
|European Classification||G06F9/52E, G06F12/08B4P2R, G06F12/08B4L|
|Jan 15, 2008||AS||Assignment|
Owner name: INTERNATIONAL BUSINESS MACHINES CORPORATION, NEW Y
Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:SU, GONG;REEL/FRAME:020364/0651
Effective date: 20070122
|Sep 17, 2012||REMI||Maintenance fee reminder mailed|
|Jan 28, 2013||FPAY||Fee payment|
Year of fee payment: 4
|Jan 28, 2013||SULP||Surcharge for late payment|
|Jul 18, 2016||FPAY||Fee payment|
Year of fee payment: 8